Electrical load forecasting for V2G scheduling: A feature-driven multi-head attention-LSTM approach

Electrical load forecasting (ELF) is a critical technology for vehicle-to-grid (V2G) scheduling, as it provides the necessary information to achieve the scheduling objective of minimizing load variance. Existing studies have demonstrated that the closer the relative magnitude relationships among forecasted loads at different time periods are to those of the actual loads, the more satisfactory V2G scheduling performance can be achieved. Inspired by the above conclusion, this paper proposes an ELF approach for V2G scheduling. By calculating the difference between load values at each time period and the daily average, a feature called load relative magnitude (LRM) is constructed to provide relative magnitude relationships between loads for the forecasting model, which is beneficial for enhancing V2G scheduling performance. Moreover, this approach builds a feature-driven multi-head attention-long short-term memory (FDMHA-LSTM) model and the constructed LRM feature is the driver. In particular, given that the multi-head attention (MHA) mechanism can be guided to focus on the key parts of the task, the LRM feature is employed to weight the Key matrix for enhancing prediction accuracy during peak and valley periods, and these periods are particularly important for V2G scheduling performance. Furthermore, extensive experiments on real-world load data demonstrate the effectiveness and superiority of the proposed model. Specifically, the proposed model can boost V2G scheduling performance by 15% to 22.3% under scenarios with varying numbers of EVs compared to LSTM model, and also demonstrates advantages over CNN and compact Transformer baselines.